mm/userfaultfd.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  mm/userfaultfd.c
   4  *
   5  *  Copyright (C) 2015  Red Hat, Inc.
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched/signal.h>
  10 #include <linux/pagemap.h>
  11 #include <linux/rmap.h>
  12 #include <linux/swap.h>
  13 #include <linux/swapops.h>
  14 #include <linux/userfaultfd_k.h>
  15 #include <linux/mmu_notifier.h>
  16 #include <linux/hugetlb.h>
  17 #include <linux/shmem_fs.h>
  18 #include <asm/tlbflush.h>
  19 #include "internal.h"
  20
  21 static __always_inline
  22 struct vm_area_struct *find_dst_vma(struct mm_struct *dst_mm,
  23                                     unsigned long dst_start,
  24                                     unsigned long len)
  25 {
  26         /*
  27          * Make sure that the dst range is both valid and fully within a
  28          * single existing vma.
  29          */
  30         struct vm_area_struct *dst_vma;
  31
  32         dst_vma = find_vma(dst_mm, dst_start);
  33         if (!dst_vma)
  34                 return NULL;
  35
  36         if (dst_start < dst_vma->vm_start ||
  37             dst_start + len > dst_vma->vm_end)
  38                 return NULL;
  39
  40         /*
  41          * Check the vma is registered in uffd, this is required to
  42          * enforce the VM_MAYWRITE check done at uffd registration
  43          * time.
  44          */
  45         if (!dst_vma->vm_userfaultfd_ctx.ctx)
  46                 return NULL;
  47
  48         return dst_vma;
  49 }
  50
  51 /*
  52  * Install PTEs, to map dst_addr (within dst_vma) to page.
  53  *
  54  * This function handles both MCOPY_ATOMIC_NORMAL and _CONTINUE for both shmem
  55  * and anon, and for both shared and private VMAs.
  56  */
  57 int mfill_atomic_install_pte(struct mm_struct *dst_mm, pmd_t *dst_pmd,
  58                              struct vm_area_struct *dst_vma,
  59                              unsigned long dst_addr, struct page *page,
  60                              bool newly_allocated, bool wp_copy)
  61 {
  62         int ret;
  63         pte_t _dst_pte, *dst_pte;
  64         bool writable = dst_vma->vm_flags & VM_WRITE;
  65         bool vm_shared = dst_vma->vm_flags & VM_SHARED;
  66         bool page_in_cache = page->mapping;
  67         spinlock_t *ptl;
  68         struct inode *inode;
  69         pgoff_t offset, max_off;
  70
  71         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
  72         _dst_pte = pte_mkdirty(_dst_pte);
  73         if (page_in_cache && !vm_shared)
  74                 writable = false;
  75         if (writable) {
  76                 if (wp_copy)
  77                         _dst_pte = pte_mkuffd_wp(_dst_pte);
  78                 else
  79                         _dst_pte = pte_mkwrite(_dst_pte);
  80         }
  81
  82         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
  83
  84         if (vma_is_shmem(dst_vma)) {
  85                 /* serialize against truncate with the page table lock */
  86                 inode = dst_vma->vm_file->f_inode;
  87                 offset = linear_page_index(dst_vma, dst_addr);
  88                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
  89                 ret = -EFAULT;
  90                 if (unlikely(offset >= max_off))
  91                         goto out_unlock;
  92         }
  93
  94         ret = -EEXIST;
  95         if (!pte_none(*dst_pte))
  96                 goto out_unlock;
  97
  98         if (page_in_cache) {
  99                 /* Usually, cache pages are already added to LRU */
 100                 if (newly_allocated)
 101                         lru_cache_add(page);
 102                 page_add_file_rmap(page, dst_vma, false);
 103         } else {
 104                 page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
 105                 lru_cache_add_inactive_or_unevictable(page, dst_vma);
 106         }
 107
 108         /*
 109          * Must happen after rmap, as mm_counter() checks mapping (via
 110          * PageAnon()), which is set by __page_set_anon_rmap().
 111          */
 112         inc_mm_counter(dst_mm, mm_counter(page));
 113
 114         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 115
 116         /* No need to invalidate - it was non-present before */
 117         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 118         ret = 0;
 119 out_unlock:
 120         pte_unmap_unlock(dst_pte, ptl);
 121         return ret;
 122 }
 123
 124 static int mcopy_atomic_pte(struct mm_struct *dst_mm,
 125                             pmd_t *dst_pmd,
 126                             struct vm_area_struct *dst_vma,
 127                             unsigned long dst_addr,
 128                             unsigned long src_addr,
 129                             struct page **pagep,
 130                             bool wp_copy)
 131 {
 132         void *page_kaddr;
 133         int ret;
 134         struct page *page;
 135
 136         if (!*pagep) {
 137                 ret = -ENOMEM;
 138                 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
 139                 if (!page)
 140                         goto out;
 141
 142                 page_kaddr = kmap_atomic(page);
 143                 ret = copy_from_user(page_kaddr,
 144                                      (const void __user *) src_addr,
 145                                      PAGE_SIZE);
 146                 kunmap_atomic(page_kaddr);
 147
 148                 /* fallback to copy_from_user outside mmap_lock */
 149                 if (unlikely(ret)) {
 150                         ret = -ENOENT;
 151                         *pagep = page;
 152                         /* don't free the page */
 153                         goto out;
 154                 }
 155
 156                 flush_dcache_page(page);
 157         } else {
 158                 page = *pagep;
 159                 *pagep = NULL;
 160         }
 161
 162         /*
 163          * The memory barrier inside __SetPageUptodate makes sure that
 164          * preceding stores to the page contents become visible before
 165          * the set_pte_at() write.
 166          */
 167         __SetPageUptodate(page);
 168
 169         ret = -ENOMEM;
 170         if (mem_cgroup_charge(page_folio(page), dst_mm, GFP_KERNEL))
 171                 goto out_release;
 172
 173         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 174                                        page, true, wp_copy);
 175         if (ret)
 176                 goto out_release;
 177 out:
 178         return ret;
 179 out_release:
 180         put_page(page);
 181         goto out;
 182 }
 183
 184 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
 185                               pmd_t *dst_pmd,
 186                               struct vm_area_struct *dst_vma,
 187                               unsigned long dst_addr)
 188 {
 189         pte_t _dst_pte, *dst_pte;
 190         spinlock_t *ptl;
 191         int ret;
 192         pgoff_t offset, max_off;
 193         struct inode *inode;
 194
 195         _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
 196                                          dst_vma->vm_page_prot));
 197         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
 198         if (dst_vma->vm_file) {
 199                 /* the shmem MAP_PRIVATE case requires checking the i_size */
 200                 inode = dst_vma->vm_file->f_inode;
 201                 offset = linear_page_index(dst_vma, dst_addr);
 202                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 203                 ret = -EFAULT;
 204                 if (unlikely(offset >= max_off))
 205                         goto out_unlock;
 206         }
 207         ret = -EEXIST;
 208         if (!pte_none(*dst_pte))
 209                 goto out_unlock;
 210         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 211         /* No need to invalidate - it was non-present before */
 212         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 213         ret = 0;
 214 out_unlock:
 215         pte_unmap_unlock(dst_pte, ptl);
 216         return ret;
 217 }
 218
 219 /* Handles UFFDIO_CONTINUE for all shmem VMAs (shared or private). */
 220 static int mcontinue_atomic_pte(struct mm_struct *dst_mm,
 221                                 pmd_t *dst_pmd,
 222                                 struct vm_area_struct *dst_vma,
 223                                 unsigned long dst_addr,
 224                                 bool wp_copy)
 225 {
 226         struct inode *inode = file_inode(dst_vma->vm_file);
 227         pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
 228         struct page *page;
 229         int ret;
 230
 231         ret = shmem_getpage(inode, pgoff, &page, SGP_READ);
 232         if (ret)
 233                 goto out;
 234         if (!page) {
 235                 ret = -EFAULT;
 236                 goto out;
 237         }
 238
 239         if (PageHWPoison(page)) {
 240                 ret = -EIO;
 241                 goto out_release;
 242         }
 243
 244         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 245                                        page, false, wp_copy);
 246         if (ret)
 247                 goto out_release;
 248
 249         unlock_page(page);
 250         ret = 0;
 251 out:
 252         return ret;
 253 out_release:
 254         unlock_page(page);
 255         put_page(page);
 256         goto out;
 257 }
 258
 259 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
 260 {
 261         pgd_t *pgd;
 262         p4d_t *p4d;
 263         pud_t *pud;
 264
 265         pgd = pgd_offset(mm, address);
 266         p4d = p4d_alloc(mm, pgd, address);
 267         if (!p4d)
 268                 return NULL;
 269         pud = pud_alloc(mm, p4d, address);
 270         if (!pud)
 271                 return NULL;
 272         /*
 273          * Note that we didn't run this because the pmd was
 274          * missing, the *pmd may be already established and in
 275          * turn it may also be a trans_huge_pmd.
 276          */
 277         return pmd_alloc(mm, pud, address);
 278 }
 279
 280 #ifdef CONFIG_HUGETLB_PAGE
 281 /*
 282  * __mcopy_atomic processing for HUGETLB vmas.  Note that this routine is
 283  * called with mmap_lock held, it will release mmap_lock before returning.
 284  */
 285 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 286                                               struct vm_area_struct *dst_vma,
 287                                               unsigned long dst_start,
 288                                               unsigned long src_start,
 289                                               unsigned long len,
 290                                               enum mcopy_atomic_mode mode)
 291 {
 292         int vm_shared = dst_vma->vm_flags & VM_SHARED;
 293         ssize_t err;
 294         pte_t *dst_pte;
 295         unsigned long src_addr, dst_addr;
 296         long copied;
 297         struct page *page;
 298         unsigned long vma_hpagesize;
 299         pgoff_t idx;
 300         u32 hash;
 301         struct address_space *mapping;
 302
 303         /*
 304          * There is no default zero huge page for all huge page sizes as
 305          * supported by hugetlb.  A PMD_SIZE huge pages may exist as used
 306          * by THP.  Since we can not reliably insert a zero page, this
 307          * feature is not supported.
 308          */
 309         if (mode == MCOPY_ATOMIC_ZEROPAGE) {
 310                 mmap_read_unlock(dst_mm);
 311                 return -EINVAL;
 312         }
 313
 314         src_addr = src_start;
 315         dst_addr = dst_start;
 316         copied = 0;
 317         page = NULL;
 318         vma_hpagesize = vma_kernel_pagesize(dst_vma);
 319
 320         /*
 321          * Validate alignment based on huge page size
 322          */
 323         err = -EINVAL;
 324         if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
 325                 goto out_unlock;
 326
 327 retry:
 328         /*
 329          * On routine entry dst_vma is set.  If we had to drop mmap_lock and
 330          * retry, dst_vma will be set to NULL and we must lookup again.
 331          */
 332         if (!dst_vma) {
 333                 err = -ENOENT;
 334                 dst_vma = find_dst_vma(dst_mm, dst_start, len);
 335                 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
 336                         goto out_unlock;
 337
 338                 err = -EINVAL;
 339                 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
 340                         goto out_unlock;
 341
 342                 vm_shared = dst_vma->vm_flags & VM_SHARED;
 343         }
 344
 345         /*
 346          * If not shared, ensure the dst_vma has a anon_vma.
 347          */
 348         err = -ENOMEM;
 349         if (!vm_shared) {
 350                 if (unlikely(anon_vma_prepare(dst_vma)))
 351                         goto out_unlock;
 352         }
 353
 354         while (src_addr < src_start + len) {
 355                 BUG_ON(dst_addr >= dst_start + len);
 356
 357                 /*
 358                  * Serialize via i_mmap_rwsem and hugetlb_fault_mutex.
 359                  * i_mmap_rwsem ensures the dst_pte remains valid even
 360                  * in the case of shared pmds.  fault mutex prevents
 361                  * races with other faulting threads.
 362                  */
 363                 mapping = dst_vma->vm_file->f_mapping;
 364                 i_mmap_lock_read(mapping);
 365                 idx = linear_page_index(dst_vma, dst_addr);
 366                 hash = hugetlb_fault_mutex_hash(mapping, idx);
 367                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 368
 369                 err = -ENOMEM;
 370                 dst_pte = huge_pte_alloc(dst_mm, dst_vma, dst_addr, vma_hpagesize);
 371                 if (!dst_pte) {
 372                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 373                         i_mmap_unlock_read(mapping);
 374                         goto out_unlock;
 375                 }
 376
 377                 if (mode != MCOPY_ATOMIC_CONTINUE &&
 378                     !huge_pte_none(huge_ptep_get(dst_pte))) {
 379                         err = -EEXIST;
 380                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 381                         i_mmap_unlock_read(mapping);
 382                         goto out_unlock;
 383                 }
 384
 385                 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
 386                                                dst_addr, src_addr, mode, &page);
 387
 388                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 389                 i_mmap_unlock_read(mapping);
 390
 391                 cond_resched();
 392
 393                 if (unlikely(err == -ENOENT)) {
 394                         mmap_read_unlock(dst_mm);
 395                         BUG_ON(!page);
 396
 397                         err = copy_huge_page_from_user(page,
 398                                                 (const void __user *)src_addr,
 399                                                 vma_hpagesize / PAGE_SIZE,
 400                                                 true);
 401                         if (unlikely(err)) {
 402                                 err = -EFAULT;
 403                                 goto out;
 404                         }
 405                         mmap_read_lock(dst_mm);
 406
 407                         dst_vma = NULL;
 408                         goto retry;
 409                 } else
 410                         BUG_ON(page);
 411
 412                 if (!err) {
 413                         dst_addr += vma_hpagesize;
 414                         src_addr += vma_hpagesize;
 415                         copied += vma_hpagesize;
 416
 417                         if (fatal_signal_pending(current))
 418                                 err = -EINTR;
 419                 }
 420                 if (err)
 421                         break;
 422         }
 423
 424 out_unlock:
 425         mmap_read_unlock(dst_mm);
 426 out:
 427         if (page)
 428                 put_page(page);
 429         BUG_ON(copied < 0);
 430         BUG_ON(err > 0);
 431         BUG_ON(!copied && !err);
 432         return copied ? copied : err;
 433 }
 434 #else /* !CONFIG_HUGETLB_PAGE */
 435 /* fail at build time if gcc attempts to use this */
 436 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 437                                       struct vm_area_struct *dst_vma,
 438                                       unsigned long dst_start,
 439                                       unsigned long src_start,
 440                                       unsigned long len,
 441                                       enum mcopy_atomic_mode mode);
 442 #endif /* CONFIG_HUGETLB_PAGE */
 443
 444 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
 445                                                 pmd_t *dst_pmd,
 446                                                 struct vm_area_struct *dst_vma,
 447                                                 unsigned long dst_addr,
 448                                                 unsigned long src_addr,
 449                                                 struct page **page,
 450                                                 enum mcopy_atomic_mode mode,
 451                                                 bool wp_copy)
 452 {
 453         ssize_t err;
 454
 455         if (mode == MCOPY_ATOMIC_CONTINUE) {
 456                 return mcontinue_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 457                                             wp_copy);
 458         }
 459
 460         /*
 461          * The normal page fault path for a shmem will invoke the
 462          * fault, fill the hole in the file and COW it right away. The
 463          * result generates plain anonymous memory. So when we are
 464          * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
 465          * generate anonymous memory directly without actually filling
 466          * the hole. For the MAP_PRIVATE case the robustness check
 467          * only happens in the pagetable (to verify it's still none)
 468          * and not in the radix tree.
 469          */
 470         if (!(dst_vma->vm_flags & VM_SHARED)) {
 471                 if (mode == MCOPY_ATOMIC_NORMAL)
 472                         err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
 473                                                dst_addr, src_addr, page,
 474                                                wp_copy);
 475                 else
 476                         err = mfill_zeropage_pte(dst_mm, dst_pmd,
 477                                                  dst_vma, dst_addr);
 478         } else {
 479                 VM_WARN_ON_ONCE(wp_copy);
 480                 err = shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
 481                                              dst_addr, src_addr,
 482                                              mode != MCOPY_ATOMIC_NORMAL,
 483                                              page);
 484         }
 485
 486         return err;
 487 }
 488
 489 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
 490                                               unsigned long dst_start,
 491                                               unsigned long src_start,
 492                                               unsigned long len,
 493                                               enum mcopy_atomic_mode mcopy_mode,
 494                                               atomic_t *mmap_changing,
 495                                               __u64 mode)
 496 {
 497         struct vm_area_struct *dst_vma;
 498         ssize_t err;
 499         pmd_t *dst_pmd;
 500         unsigned long src_addr, dst_addr;
 501         long copied;
 502         struct page *page;
 503         bool wp_copy;
 504
 505         /*
 506          * Sanitize the command parameters:
 507          */
 508         BUG_ON(dst_start & ~PAGE_MASK);
 509         BUG_ON(len & ~PAGE_MASK);
 510
 511         /* Does the address range wrap, or is the span zero-sized? */
 512         BUG_ON(src_start + len <= src_start);
 513         BUG_ON(dst_start + len <= dst_start);
 514
 515         src_addr = src_start;
 516         dst_addr = dst_start;
 517         copied = 0;
 518         page = NULL;
 519 retry:
 520         mmap_read_lock(dst_mm);
 521
 522         /*
 523          * If memory mappings are changing because of non-cooperative
 524          * operation (e.g. mremap) running in parallel, bail out and
 525          * request the user to retry later
 526          */
 527         err = -EAGAIN;
 528         if (mmap_changing && atomic_read(mmap_changing))
 529                 goto out_unlock;
 530
 531         /*
 532          * Make sure the vma is not shared, that the dst range is
 533          * both valid and fully within a single existing vma.
 534          */
 535         err = -ENOENT;
 536         dst_vma = find_dst_vma(dst_mm, dst_start, len);
 537         if (!dst_vma)
 538                 goto out_unlock;
 539
 540         err = -EINVAL;
 541         /*
 542          * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
 543          * it will overwrite vm_ops, so vma_is_anonymous must return false.
 544          */
 545         if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
 546             dst_vma->vm_flags & VM_SHARED))
 547                 goto out_unlock;
 548
 549         /*
 550          * validate 'mode' now that we know the dst_vma: don't allow
 551          * a wrprotect copy if the userfaultfd didn't register as WP.
 552          */
 553         wp_copy = mode & UFFDIO_COPY_MODE_WP;
 554         if (wp_copy && !(dst_vma->vm_flags & VM_UFFD_WP))
 555                 goto out_unlock;
 556
 557         /*
 558          * If this is a HUGETLB vma, pass off to appropriate routine
 559          */
 560         if (is_vm_hugetlb_page(dst_vma))
 561                 return  __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
 562                                                 src_start, len, mcopy_mode);
 563
 564         if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
 565                 goto out_unlock;
 566         if (!vma_is_shmem(dst_vma) && mcopy_mode == MCOPY_ATOMIC_CONTINUE)
 567                 goto out_unlock;
 568
 569         /*
 570          * Ensure the dst_vma has a anon_vma or this page
 571          * would get a NULL anon_vma when moved in the
 572          * dst_vma.
 573          */
 574         err = -ENOMEM;
 575         if (!(dst_vma->vm_flags & VM_SHARED) &&
 576             unlikely(anon_vma_prepare(dst_vma)))
 577                 goto out_unlock;
 578
 579         while (src_addr < src_start + len) {
 580                 pmd_t dst_pmdval;
 581
 582                 BUG_ON(dst_addr >= dst_start + len);
 583
 584                 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 585                 if (unlikely(!dst_pmd)) {
 586                         err = -ENOMEM;
 587                         break;
 588                 }
 589
 590                 dst_pmdval = pmd_read_atomic(dst_pmd);
 591                 /*
 592                  * If the dst_pmd is mapped as THP don't
 593                  * override it and just be strict.
 594                  */
 595                 if (unlikely(pmd_trans_huge(dst_pmdval))) {
 596                         err = -EEXIST;
 597                         break;
 598                 }
 599                 if (unlikely(pmd_none(dst_pmdval)) &&
 600                     unlikely(__pte_alloc(dst_mm, dst_pmd))) {
 601                         err = -ENOMEM;
 602                         break;
 603                 }
 604                 /* If an huge pmd materialized from under us fail */
 605                 if (unlikely(pmd_trans_huge(*dst_pmd))) {
 606                         err = -EFAULT;
 607                         break;
 608                 }
 609
 610                 BUG_ON(pmd_none(*dst_pmd));
 611                 BUG_ON(pmd_trans_huge(*dst_pmd));
 612
 613                 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 614                                        src_addr, &page, mcopy_mode, wp_copy);
 615                 cond_resched();
 616
 617                 if (unlikely(err == -ENOENT)) {
 618                         void *page_kaddr;
 619
 620                         mmap_read_unlock(dst_mm);
 621                         BUG_ON(!page);
 622
 623                         page_kaddr = kmap(page);
 624                         err = copy_from_user(page_kaddr,
 625                                              (const void __user *) src_addr,
 626                                              PAGE_SIZE);
 627                         kunmap(page);
 628                         if (unlikely(err)) {
 629                                 err = -EFAULT;
 630                                 goto out;
 631                         }
 632                         flush_dcache_page(page);
 633                         goto retry;
 634                 } else
 635                         BUG_ON(page);
 636
 637                 if (!err) {
 638                         dst_addr += PAGE_SIZE;
 639                         src_addr += PAGE_SIZE;
 640                         copied += PAGE_SIZE;
 641
 642                         if (fatal_signal_pending(current))
 643                                 err = -EINTR;
 644                 }
 645                 if (err)
 646                         break;
 647         }
 648
 649 out_unlock:
 650         mmap_read_unlock(dst_mm);
 651 out:
 652         if (page)
 653                 put_page(page);
 654         BUG_ON(copied < 0);
 655         BUG_ON(err > 0);
 656         BUG_ON(!copied && !err);
 657         return copied ? copied : err;
 658 }
 659
 660 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
 661                      unsigned long src_start, unsigned long len,
 662                      atomic_t *mmap_changing, __u64 mode)
 663 {
 664         return __mcopy_atomic(dst_mm, dst_start, src_start, len,
 665                               MCOPY_ATOMIC_NORMAL, mmap_changing, mode);
 666 }
 667
 668 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
 669                        unsigned long len, atomic_t *mmap_changing)
 670 {
 671         return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_ZEROPAGE,
 672                               mmap_changing, 0);
 673 }
 674
 675 ssize_t mcopy_continue(struct mm_struct *dst_mm, unsigned long start,
 676                        unsigned long len, atomic_t *mmap_changing)
 677 {
 678         return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_CONTINUE,
 679                               mmap_changing, 0);
 680 }
 681
 682 int mwriteprotect_range(struct mm_struct *dst_mm, unsigned long start,
 683                         unsigned long len, bool enable_wp,
 684                         atomic_t *mmap_changing)
 685 {
 686         struct vm_area_struct *dst_vma;
 687         pgprot_t newprot;
 688         int err;
 689
 690         /*
 691          * Sanitize the command parameters:
 692          */
 693         BUG_ON(start & ~PAGE_MASK);
 694         BUG_ON(len & ~PAGE_MASK);
 695
 696         /* Does the address range wrap, or is the span zero-sized? */
 697         BUG_ON(start + len <= start);
 698
 699         mmap_read_lock(dst_mm);
 700
 701         /*
 702          * If memory mappings are changing because of non-cooperative
 703          * operation (e.g. mremap) running in parallel, bail out and
 704          * request the user to retry later
 705          */
 706         err = -EAGAIN;
 707         if (mmap_changing && atomic_read(mmap_changing))
 708                 goto out_unlock;
 709
 710         err = -ENOENT;
 711         dst_vma = find_dst_vma(dst_mm, start, len);
 712         /*
 713          * Make sure the vma is not shared, that the dst range is
 714          * both valid and fully within a single existing vma.
 715          */
 716         if (!dst_vma || (dst_vma->vm_flags & VM_SHARED))
 717                 goto out_unlock;
 718         if (!userfaultfd_wp(dst_vma))
 719                 goto out_unlock;
 720         if (!vma_is_anonymous(dst_vma))
 721                 goto out_unlock;
 722
 723         if (enable_wp)
 724                 newprot = vm_get_page_prot(dst_vma->vm_flags & ~(VM_WRITE));
 725         else
 726                 newprot = vm_get_page_prot(dst_vma->vm_flags);
 727
 728         change_protection(dst_vma, start, start + len, newprot,
 729                           enable_wp ? MM_CP_UFFD_WP : MM_CP_UFFD_WP_RESOLVE);
 730
 731         err = 0;
 732 out_unlock:
 733         mmap_read_unlock(dst_mm);
 734         return err;
 735 }